Download neural_network_0219

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Data (Star Trek) wikipedia , lookup

Time series wikipedia , lookup

Pattern recognition wikipedia , lookup

Catastrophic interference wikipedia , lookup

Convolutional neural network wikipedia , lookup

Transcript
Information Fusion
Yu Cai
Research Article
• “Comparative Analysis of Some Neural
Network Architectures for Data Fusion”,
• Authors: Juan Cires, PA Romo, PJ Zufiria,
• IEEE International Conference on Neural
Networks, 1995
Abstract
• The various characteristics of fusion
algorithms yield different design
alternatives for the architecture of the neural
network.
• These alternatives are summarized with
comparative results.
• This paper validate the use of neural
network for data fusion and provide a
design framework for future work
Introduction
• A data fusion system combines the
information from several sensors, or several
sensor information processing modules to
reduce the uncertainty of the information or
to produce information that is not available
from any of the sensors by themselves.
Classification of data fusion
• By the type of information to be fused
– Congruent information data fusion: from same type
sensors or a sensor over a period of time
– Complementary information data fusion: from different
type sensors
• By the level of abstraction
– Like images: signal, pixel, feature or symbolic level
– Centralized architecture: low level fusion with low
level data input
– Distributed architecture: sensor fusion locally, output
high level data for further fusion
Classification of data fusion
• By the interaction of fusion modules
– Strongly coupled: modules output depends on
each other
– Loosely coupled: independent modules with
little/no interaction
• By functional point of view
– Positional fusion: the position/state of the
observed object.
– Identity fusion: the identity of those object.
Neural network
• An artificial neural network can be defined as a
set of processing elements (neurons), a specific
topology of weighted interconnection between
these elements, and a learning law which updates
the connection weights.
• Neurons provide non-linear input/output transfer
functions
• Neural network topology fit into: a feed forward
topology and a recursive topology
• Learning law includes supervised learning,
unsupervised learning and reinforcement learning.
Why neural network for data fusion
• Adaptive fusion inference:
– Neural network can infer the relationship
between the fusion output and the multiple
inputs
• Incomplete information generalization
– Information is noisy, distorted and incomplete
• Non-linear filtering of noise
• Parallel data computing
Neural network for data fusion
• The architecture of neural network reflects
the different characteristics of fusion
algorithms, and the types of relations
between modules.
– Different types of information => different
inputs to neural network
– Level of data fusion =>where to use neural
network
– The coupling alternatives => interconnection of
the neural networks
Simulation
• Simulation Environment
– Video image sensor: 25*25 pixel
– Ultrasound sensor: 32*1 pixel
– Objects on table: sphere, block, others
• Goal: Train neural networks for using
sensor data to estimate the object position
(center of mass).
Image by sensor
System
System
• Neural network A, B
• Loosely coupled system C vs. Strongly coupled
system D
• After get A and B, the types of C:
– C-NLC: C is a neural network, and output non linear
combination of A and B
– C-Retrain: the whole system ABC is further retrained
– C-Avg: average A and B
– C-OLC: get an optimal linear combination of A and B
by minimizing the mean squared error
– C-N-OLC: compute weights for this linear combination
using neural network
Result
Discussion
• B is better than A because of high resolution
of ultrasound than imaging (32*1 vs 25*25)
• Loosely coupled C is better than strongly
coupled D
• For blocks, all the loosely get similar result;
but for sphere, c-retrain is the best.
Result 2
Discussion 2
• A and B is not fully trained with high error
• C-retrain still performs the best.
Conclusion
• It is possible to perform data fusion with neural
network, without knowledge of the characteristic
input signal.
• Neural network perform well in the presence of
noise.
• The result show a modular, loosely coupled
architecture perform better than a monolithic,
strongly coupled architecture.
• Within the loosely coupled, C-retrain seems to be
the best